Saurabh Bipin Chandra

Machine learning scientist · Mountain View, CA · srbhchandra@gmail.com · LinkedIn · GitHub

Principal machine learning scientist with an MS in CS from UCSD and 9+ years applying deep learning to computer vision and NLP problems. First ML hire at Gradescope (acquired by Turnitin), with experience spanning ML engineering, data science, and technical leadership — 12+ models across 10+ projects taken from research to production, most of them in settings that demand very high precision. Currently tech-leading AI content detection, with a focus on agentic data generation and automated training and evaluation pipelines. Previously four years on optimized CV algorithms for automotive ADAS and multimedia frameworks for embedded systems.

Experience

Turnitin — Principal Machine Learning Scientist 2016; 2017 – present

AI-generated text detection

• Tech-leading the AI-generated text detection system; reduced release cadence from 4–6 months to 1 month.
• Trained, evaluated, and deployed a 3-model ensemble detecting AI-written, AI-paraphrased, and AI-bypassed ("humanized") content.
• State-of-the-art performance among competitors while holding a false-positive rate below 0.5% across all groups, evaluated on 1M+ documents; ran extensive statistical testing for fairness and absence of bias across groups.
• Spearheaded AI-paraphrased text detection from conception to deployment, tech-leading 3 AI team members and coordinating a cross-functional team of 7 across design, PM, integrations, and QA — including a three-phase incremental release strategy.
• Building automated, rerunnable training and evaluation pipelines covering the full model lifecycle from data preparation to production deployment.

Clarity — student writing assistant & instructor report

• Co-led development, evaluation, and deployment of the LLM-based student writing assistant, providing real-time feedback and Socratic guidance. Product named a TIME Best Invention of 2025.
• Sole developer of the AI-powered instructor report's three LLM components: a prompt classifier categorizing student requests by writing stage and intent, plus two summarization systems (full-conversation and per-category).
• Cut LLM cost 81–89% while maintaining quality via multi-region model strategies across US/EU/APAC.
• Built reusable LLM infrastructure adopted organization-wide: an enhanced GEval evaluation framework, a prompt builder for systematic prompt development, evaluation tracking (Langfuse, Weave), and a multilingual evaluation framework.
• Developed a student agent for systematic multi-turn testing against diverse personas, surfacing edge cases and enabling multilingual expansion.

Gradescope — document AI & grading automation

• ML/CV technical lead for the paper-to-digital automated grading workflow, which cuts grading time by ~50%. Built and maintained all ML/CV services on a two-person team, including data flywheels (discovery, tagging, annotation, production evaluation) and one-script on-premise training.
• Handwritten text recognition — CNN + BiLSTM network for single-line text, with a CV solution splitting multi-line text into lines.
• Automated scan splitting — split multi-page PDFs into individual student submissions; 100% accuracy at 80% recall through confidence tuning, enabling a workflow needing no manual review when confident.
• Image alignment, masking, and overlaying — projective transformation aligning submissions to templates, with a novel iterative selective feature-outlier removal method and a custom alignment metric.
• Automated roster matching — deep network matching inferred handwritten names/IDs to roster entries only when fully confident; 100% precision in production via a novel "not-confident" softmax label, paired with a three-tier confidence-based user workflow.
• Multiple-choice grading — inception-based fully convolutional detector for mark regions with a confidence network and contour analysis, plus a model detecting student ink within marks.
• Blank answer detection — Siamese-like residual CNN robust to misaligned and incorrectly overlaid images.
• Question type classification — CNN classifier for question type (multiple choice, fill-in-the-blank, short answer) and answer type.
• Fast multi-head attention — optimized forward pass reducing time complexity from O(T²) to O(T), significantly speeding multi-line HTR and enabling low-cost deployment. Acknowledged in the ICDAR 2021 paper "Full Page Handwriting Recognition via Image to Sequence Extraction."
• Template PDF parser — graph-based solution parsing PDF structure and segmenting regions, text, and images.
• Assisted the NSF SBIR project "Grading 10x Faster with AI Assistance," demonstrating few-shot feasibility for grading unseen questions after an instructor grades ≤10% of answers.

Texas Instruments India — Senior Software Engineer 2011 – 2015

• Optimized machine vision algorithms for real-time, safety-critical automotive driver assistance (ADAS) on EVE, a SIMD compute engine. The lens distortion correction algorithm was 4× better than the existing DSP solution and was productized for surround view.
• Sole developer of an OS-agnostic remote multimedia framework stack (Android, Linux, QNX) for TI's heterogeneous OMAP4/5 processors, offloading multimedia work to a remote core over IPC with resource, memory, and power management.
• Built an OpenMAX-based pipelined scalable video codec encoder using two SIMD compute engines alongside the hardware accelerator to deliver 1080p30.

Collaborations

• Collaborated with Dr. Nikita Hari (University of Cambridge) on her PhD, "Gallium Nitride (GaN) Power Electronics using Machine Learning" — developed a deep network predicting time-series voltage waveforms for GaN devices from intrinsic properties and device current.
• Created labs for the Full Stack Deep Learning bootcamp (2018, 2020, 2022), now offered as a course at UC Berkeley — a CNN text-line region predictor and ResNet + Transformer handwriting recognition models for single lines and paragraphs.

Skills

• Languages — Python, SQL (PostgreSQL), C
• ML & data science — PyTorch, TensorFlow, Keras, scikit-learn, NumPy, pandas, SciPy, Wandb, Matplotlib
• CV & NLP — OpenCV, Hugging Face, spaCy
• LLM & prompt engineering — LiteLLM, OpenAI, AWS Bedrock, Langfuse, Weave, GEval
• Cloud & deployment — AWS (Lambda, S3, SageMaker, Redshift, Bedrock), Docker
• Version control, CI/CD — Git, CircleCI, Jenkins

Education

• MS, Computer Science — University of California San Diego, 2017. AI, data mining, ML, neural networks, image recognition. GPA 3.77/4.
• B.Tech, Computer Engineering — National Institute of Technology Karnataka, 2011. GPA 8.9/10.